Instructions to use akaclinicalco/edgecdss-v3-research with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use akaclinicalco/edgecdss-v3-research with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf akaclinicalco/edgecdss-v3-research
# Run inference directly in the terminal:
llama cli -hf akaclinicalco/edgecdss-v3-research

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf akaclinicalco/edgecdss-v3-research
# Run inference directly in the terminal:
llama cli -hf akaclinicalco/edgecdss-v3-research

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf akaclinicalco/edgecdss-v3-research
# Run inference directly in the terminal:
./llama-cli -hf akaclinicalco/edgecdss-v3-research

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf akaclinicalco/edgecdss-v3-research
# Run inference directly in the terminal:
./build/bin/llama-cli -hf akaclinicalco/edgecdss-v3-research

Use Docker

docker model run hf.co/akaclinicalco/edgecdss-v3-research

LM Studio
Jan

vLLM

How to use akaclinicalco/edgecdss-v3-research with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "akaclinicalco/edgecdss-v3-research"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "akaclinicalco/edgecdss-v3-research",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/akaclinicalco/edgecdss-v3-research

Ollama
How to use akaclinicalco/edgecdss-v3-research with Ollama:
```
ollama run hf.co/akaclinicalco/edgecdss-v3-research
```
Unsloth Desktop

How to use akaclinicalco/edgecdss-v3-research with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf akaclinicalco/edgecdss-v3-research

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "akaclinicalco/edgecdss-v3-research"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use akaclinicalco/edgecdss-v3-research with Docker Model Runner:
```
docker model run hf.co/akaclinicalco/edgecdss-v3-research
```

Lemonade

How to use akaclinicalco/edgecdss-v3-research with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull akaclinicalco/edgecdss-v3-research

Run and chat with the model

lemonade run user.edgecdss-v3-research-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use akaclinicalco/edgecdss-v3-research with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf akaclinicalco/edgecdss-v3-research

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default akaclinicalco/edgecdss-v3-research

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use akaclinicalco/edgecdss-v3-research with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf akaclinicalco/edgecdss-v3-research

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "akaclinicalco/edgecdss-v3-research" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

edgecdss-v3 — research artifact, not adopted

Built with Qwen.

A LoRA distillation of Qwen2.5-3B-Instruct trained to write EdgeCDSS field cards. The teacher was Anthropic's claude-opus-5: it wrote every training answer. OpenAI's gpt-4o-mini checked each one as EdgeCDSS's validator. No human-written answers were used.

⚠️ Not for clinical use, and not adopted by the project. On the project owner's reading of its benchmark answers (2026-10-10), this model fails: several answers it served were clinically wrong. The project's offline model remains qwen2.5:3b. This card exists so the result is on the record. It is a research artifact about distillation for edge clinical decision support, and nothing more.

Source: EdgeCDSS v5.0.0 (changelog, safety statement). Bench record: DISTILL_BENCH_edgecdss-v3-d7.md (with qwen2.5:3b as validator) and DISTILL_BENCH_edgecdss-v3.md (as its own validator).

Licence

This model is a derivative of Qwen2.5-3B-Instruct and is distributed under the Qwen RESEARCH LICENSE AGREEMENT: non-commercial use (research or evaluation) only; commercial use needs a licence from Alibaba Cloud. A copy of the agreement and a NOTICE file ship with these weights. The EdgeCDSS repository's own code is MIT; that does not apply to these weights.

What it is


Base	`Qwen/Qwen2.5-3B-Instruct` (trained on `mlx-community/Qwen2.5-3B-Instruct-4bit`)
Method	LoRA (`mlx_lm.lora`): rank 8, scale 20, dropout 0, 16 layers; 300 iterations, learning rate 1e-4, batch 1, gradient checkpointing, max sequence length 6656, seed 0
Loss	train 2.071 → 0.366; validation 2.880 → 0.625 (iteration 300)
Artifact	fused, then GGUF Q4_K_M (1.93 GB), sha256 `a5a24a6442d076d4e580c4e686a10b375973f2d705f9baf781643246a53bb8c4`; Ollama tag `edgecdss-v3` (93a43d964fff) with the base model's TEMPLATE and SYSTEM
Teacher	`claude-opus-5` (Anthropic) wrote every training answer, under the EdgeCDSS system prompt; `gpt-4o-mini` (OpenAI) validated each one, with EdgeCDSS's deterministic checks
Data	221 rows (205 train / 16 validation) from 255 field questions; 34 dropped (20 named a drug the question did not ask about, 8 changed by the pipeline, 6 held). Every row carries the full live system prompt. Questions overlapping the evaluation set were excluded by hash. The dataset is not released.

How it was evaluated

The EdgeCDSS distillation bench (make bench, tools/distill): the project's 29-case endpoint suite (run_tests.sh), and a signed 30-scenario evaluation set run through the full pipeline, on an NVIDIA Jetson Orin Nano (Ollama 0.34.2, 25 W).

	`qwen2.5:3b` (base)	`edgecdss-v3` as its own validator	`edgecdss-v3`, validator `qwen2.5:3b`
`run_tests.sh`	29 / 29	28 / 29	29 / 29
30-set model answers served (of 23)	15–16	0 (19 of 23 validator replies unreadable)	17
Specifics present, same served scenarios	4 / 39	—	8 / 39
Mean served answer (tokens)	87	—	234
Latency median / p95 (s)	8.1 / 13.0	28.2 / 34.6	16.4 / 23.5

Full records, as of v5.0.0: DISTILL_BENCH_edgecdss-v3-d7.md and DISTILL_BENCH_edgecdss-v3.md.

Why it was not adopted

The counts improved on the base. The answers did not hold up. Read in full, answers the model wrote and the pipeline served included:

"tourniquet low and loose" and debridement before control, for a fresh arterial bleed;
"reapply" a tourniquet on loss of distal pulse;
"Do not re-warm until professional help arrives", for a hypothermic arrest;
"Treat as a presumed cardiac arrest … chest compressions", for an adult having a seizure;
repositioning and "secure the tube", with no needle decompression, for a tension pneumothorax.

Other behaviour observed:

As a validator it is unusable: asked for a verdict, it writes a field card instead (19 of 23 calls), sometimes stating unsigned transfusion volumes.
It states doses and rules no source gave it: a "rule of two" ketamine range and a vial concentration in a bare prompt; "Draw 50 mcg of fentanyl IV (50 mg)", a thousandfold unit slip next to the correct dose.

In EdgeCDSS every such dose is held by deterministic checks; procedural advice is not, which is why the project's rule is that benchmark counts are advisory and a distilled model ships only when every served answer has been read and none is clinically wrong.

Intended use

Research on distilling a large model's clinical answers into a small on-device model: what improves (specific terms, format), what does not (clinical correctness), and what the surrounding safety layer must catch.

Out of scope

Any clinical use, any patient-facing use, any use as a validator, and any commercial use (Qwen licence).

Downloads last month: 239

GGUF

Model size

3B params

Architecture

qwen2

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for akaclinicalco/edgecdss-v3-research

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-3B-Instruct

Quantized

(322)

this model